Tech

AI Can Create Scaryly Accurate Faces With Just Your Voice

Photos are made with light, but what if portraits of people could be made with the sound of their voices? An AI is being worked on to reconstruct a person’s face with just a short recording of the person’s voice, the results are impressive and a little scary.

Artificial intelligence scientists at the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) first announced an AI algorithm called Speech2Face in a paper in 2019 and continued to refine it. hitherto.

AI can create terrifyingly accurate faces with just your voice - Photo 1.

First, the researchers designed and trained a deep neural network using millions of videos of people talking from YouTube and the Internet. During this training, the AI ​​learned the correlation between the sound of the voice and the appearance of the speaker. These correlations allow it to make the best guess about the speaker’s age, gender, and ethnicity.

Humans aren’t directly involved in the training, as researchers don’t need to manually categorize any data – the AI ​​is simply fed a large amount of video and tasked with finding connections. correlation between voice and facial features.

Once trained, the AI ​​seems to be very good at creating lifelike portraits based on voice recordings alone. This AI works better when the recordings are longer.

AI can create terrifyingly accurate faces with just your voice - Photo 2.

On the left is a real face and on the right is an AI-generated face from voice

To further analyze the accuracy of the face reconstruction, the researchers built a “face decoder” that generates a reference from the original face, ignoring irrelevant things like posture. And the light. This allows scientists to easily compare the image generated from the voice with the image of the speaker’s face.

Again, the AI ​​results are very close to real faces in most cases.

AI can create terrifyingly accurate faces with just your voice - Photo 3.

On the left is the real face, in the middle is the reference face, on the right is an image created by AI

There are some cases where the AI ​​has trouble visualizing what a speaker looks like. Factors such as accent, language, and voice pitch are the factors that cause voice-to-face mismatches, where gender, age, or ethnicity are incorrect.

People with high voices (including boys) are generally considered female while those with low voices are considered male. An Asian man speaking English results in a more un-Asian appearance than if he spoke Chinese.

AI can create terrifyingly accurate faces with just your voice - Photo 4.

AI sometimes gets wrong gender, race, age

The researchers said they had privacy and ethical considerations surrounding the project. All actual usage plans (if any) need to be carefully checked.

Law enforcement could use AI to portray a suspect if the only evidence is a voice recording. However, this can cause a lot of controversy regarding privacy and ethics.

On the other hand, it could have a negative impact on content creators on YouTube and TikTok, who are trying to protect their private lives by just voicing and not appearing in front of the camera.

While an AI that can create accurate portraits of people just from their voices is a fascinating concept and something that only exists in science fiction, that is not the goal of the researchers. They say the study aims to provide a more comprehensive view of the correlation between faces and voices and could open up new research and application opportunities.

Reference: Petapixel


https://genk.vn/ai-co-the-tao-ra-khuon-mat-chinh-xac-mot-cach-dang-so-chi-bang-giong-noi-cua-ban-20220406142712807.chn

You are reading the article AI Can Create Scaryly Accurate Faces With Just Your Voice
at Blogtuan.info – Source: genk.vn – Read the original article here

Back to top button